The effects of Facebook and Instagram on the 2020 election: A deactivation experiment

The paper investigates whether access to Facebook and Instagram affects voters’ knowledge, polarization, perceived legitimacy of the electoral process, political participation, and candidate preferences during the 2020 U.S. presidential election. Concerns about social media’s impact on democracy include heightened polarization via echo chambers, varied effects on political knowledge due to exposure to both accurate and inaccurate information, potential changes in engagement and turnout, and shifts in trust in election integrity. The context is the highly polarized 2020 election, with extensive social media discourse about misinformation, fraud, vote-by-mail, and the post-election “Stop the Steal” movement. Given Instagram’s growing role among younger voters, the study evaluates both platforms using a randomized deactivation experiment to estimate causal effects on individual-level outcomes.

The study situates itself within a growing literature on social media’s political effects. Prior work suggests social media can increase polarization via like-minded networks and algorithmic curation, though exposure to opposing views may also cause backlash. Research indicates Americans frequently consume news via social media, where both mainstream content and misinformation circulate widely; a small share of users see substantial false content. Earlier deactivation experiments measured Facebook’s effects on political outcomes in the U.S. and elsewhere, but with smaller samples and less direct measurement. Related randomized and natural experiments examine social media’s influence on knowledge, engagement, and voting, with mixed findings on turnout and participation. The paper builds on the U.S. 2020 Facebook and Instagram Election Study, which provided unique access to platform data and algorithms, and contributes by examining both Facebook and Instagram during a presidential election with large samples, a longer deactivation period, directly measured outcomes, and administrative platform data.

Design: Two parallel randomized experiments, one with Facebook as the focal platform and one with Instagram. Meta sampled U.S.-based users aged 18+ who had logged in within the past month and invited them via in-feed prompts (Aug 31–Sep 12, 2020) to participate. Participants who agreed to deactivate were offered compensation for two deactivation durations: 1 week ($25) and 6 weeks ($150). Surveys: baseline (Sep 8–21), endline (Nov 4–18), and postendline (Dec 9–23). Randomization: After baseline, participants were randomized: 27% to Deactivation (paid to avoid logging in for 6 weeks) and 73% to Control (paid to avoid logging in for 1 week). Meta deactivated accounts on Sep 23–24 and reactivated Control on Sep 30 and Deactivation on Nov 4. Participants could reactivate (forfeiting payment) but remained in the study. Samples: Primary analysis samples include users with >15 minutes/day baseline usage and who could be linked to platform data: Facebook n=19,857 (5,691 passive tracking opt-in); Instagram n=15,585 (3,822 passive tracking opt-in). Samples were weighted to be representative of U.S. focal-platform users on race, party, education, and baseline activity among >15 min/day users. Direct measures: Matched participants to state voter files (validated turnout), public campaign donation records, passive tracking of apps/websites (subset), and Meta administrative data (time spent, daily use defined by viewing ≥5 pieces of content). Compliance and usage: During the 5-week treatment period (Sep 30–Nov 3), Deactivation group’s daily usage dropped from ~90% baseline to ~15–20%, implying ~83% reduction relative to Control. Postdeactivation, Deactivation users continued to spend less time than Control (reductions of 23% for Facebook and 15% for Instagram). Substitution tracking: Passive tracking measured time on other social apps (YouTube, Twitter, Snapchat), news apps (e.g., NYT, Fox News), and the other Meta app. Surveys assessed frequency of political information consumption from multiple sources. Outcomes: Primary outcomes (8) include knowledge, affective polarization, issue polarization, perceived legitimacy, participation, Trump favorability, turnout (self-reported and validated), and Trump vote. Knowledge index is the average of standardized scores on election knowledge, news knowledge, and fact knowledge (true vs. false statements, including circulating misinformation). Polarization indices combine affect toward supporters and candidates, perceived smartness, and issue positions across eight salient topics. Estimation: Preanalysis plan (registered OSF Sep 22, 2020; updated Nov 3, 2020) prespecified outcomes, subgroups, controls, weighting, and inference using two-sided tests with sharpened FDR Q-values (Q<0.05 for significance). Primary estimation used instrumental variables: Yi = τ Di + ρ Xi + vi + εi, instrumenting compliance Di with treatment assignment Ti. Compliance Di = (ŪC − Ui)/ŪC, where Ui is share of treatment-period days used; Di=1 denotes never using the platform; τ is the local average treatment effect of never using versus Control-average use for compliers. Controls Xi were selected via lasso from baseline outcome (if available), demographics, and baseline survey variables. Balance and attrition: Groups were balanced at baseline and endline; endline completion rates were 91% (Deactivation) vs 89% (Control) on Facebook and 88% vs 86% on Instagram. Differential attrition (~2 pp) was tested using methods from Ghanem et al., Oster, and Lee bounds; evidence suggests limited bias, with Lee bounds excluding zero for participation (both platforms) and Trump vote (Facebook). Robustness: Results were stable across specifications (equal weighting, controlling for party-by-response-date, excluding lasso-selected controls, excluding multi-account users, alternative weights). Subgroup analyses examined moderators: party, baseline use, undecided status, race/ethnicity (Black/Hispanic), age, gender, education, urbanicity, swing-state residence.

• Participation: Both Facebook and Instagram deactivation significantly reduced political participation indices (Facebook: −0.167 SD; Instagram: −0.090 SD; P<0.01, Q<0.01), driven largely by reductions in online activities (posting about politics and signing online petitions). No significant effects on validated political contributions.
• Turnout: No significant effects on voter turnout; 95% CIs rule out effects larger than ~±1.6 percentage points. Validated turnout results mirror self-reports.
• Knowledge: Overall knowledge index effects are small and not statistically significant. Facebook deactivation slightly reduced knowledge by 0.033 SD (P=0.069, Q=0.190; 95% CI −0.069, 0.003). Instagram’s effect is near zero (95% CI −0.028, 0.050). Components: Facebook deactivation decreased news knowledge by 0.098 SD (P<0.01, Q<0.01) and (insignificantly after multiple-testing correction) increased fact knowledge by 0.042 SD (P=0.012, Q=0.132), consistent with reduced general news exposure but improved ability to distinguish misinformation from true statements.
• Misinformation beliefs: Facebook deactivation decreased belief in specific false statements (e.g., fraudulent ballots; Hunter Biden laptop proving bribes; “Joe Biden is a pedophile”) and increased some incorrect beliefs about non-events (e.g., Trump stopping rallies), suggesting complex effects of platform access on misinformation vs. real news.
• Trust: Deactivation reduced trust in political information from the deactivated platform (Facebook: −0.040 SD, P<0.01, Q<0.01), without affecting trust in the other platform or other news sources. Instagram deactivation similarly reduced trust in Instagram.
• Polarization: No significant effects on affective or issue polarization for either platform. Point estimates suggest small reductions in affective polarization (Facebook: −0.031 SD; Instagram: −0.030 SD), statistically indistinguishable from zero after multiple-testing correction; issue polarization effects near zero with 95% CI excluding ±0.04 SD.
• Perceived legitimacy: No significant effects on perceived legitimacy of the election; 95% CI bounds rule out effects of ±0.04 SD. Subcomponents (e.g., perceived fraud) also show small, insignificant effects.
• Candidate evaluations and vote choice: No significant effects on Trump favorability (Facebook 95% CI −0.030, 0.005; Instagram 95% CI −0.014, 0.021). Facebook deactivation reduced self-reported Trump vs. Biden vote by 0.026 units (P=0.015, Q=0.076; 95% CI −0.046, −0.005), just above the preregistered significance threshold; Instagram effect is not significant (95% CI −0.012, 0.034). The magnitude implies, for example, a 1.16 pp reduction in Trump’s two-party vote share in-sample.
• Substitution: Facebook deactivation increased time on other social apps by ~8 minutes/day and on news apps by ~1 minute/day; Instagram deactivation increased other social app time by ~8 minutes/day, with little substitution to Facebook or news apps. Self-reports show reduced frequency of obtaining political information from online news and TV/radio sources (net reduction a bit over 1 time/week; ~4% relative to a Control average of 26 times/week).
• Heterogeneity: Larger effects on perceived legitimacy and participation among above-median baseline users; Democrats showed reduced knowledge; strong Democrats increased issue polarization while strong Republicans decreased (insignificantly). Some subgroups (older, men, college graduates, rural) exhibited somewhat larger negative effects on affective polarization.

The randomized deactivation experiment addresses the central question of how access to Facebook and Instagram shapes voters’ knowledge, attitudes, and behavior during a high-salience electoral period. The findings suggest that platform access increases online political participation but has limited short-term effects on polarization, perceived legitimacy, and turnout. Facebook access appears to improve awareness of current events (news knowledge) while potentially increasing susceptibility to misinformation, aligning with the coexistence of high-quality news and low-quality content on social media. The suggestive reduction in Trump vote among Facebook deactivators, although not significant at the preregistered Q<0.05 threshold, indicates that platform exposure may have partisan consequences in close elections, possibly due to campaign effectiveness or net content slant. However, small effects on correlated measures (favorability, issue positions) caution against strong causal claims about vote choice. Precisely estimated null effects on perceived legitimacy and turnout narrow the range of plausible impacts of social media access on foundational democratic attitudes and behaviors during the study period. Overall, the results refine expectations: if social media contributes to polarization or turnout changes, effects are likely small or accumulate over longer time horizons than the 5-week deactivation window.

This study provides the largest-scale experimental evidence to date on the effects of Facebook and Instagram access during a U.S. presidential election, leveraging platform deactivation, administrative data, and passive tracking. Key contributions include: (1) documenting reductions in online political participation when access is removed; (2) disentangling components of knowledge, showing reduced news knowledge but improved discrimination of misinformation with Facebook deactivation; (3) establishing precisely estimated near-zero effects on polarization, perceived legitimacy, and turnout; and (4) presenting suggestive evidence of a shift in self-reported vote choice away from Trump with Facebook deactivation. Future research should examine longer-term deactivations, broader general equilibrium effects when many users reduce usage, post-election periods when legitimacy narratives evolve, platform-specific content dynamics, and heterogeneous effects across demographics and baseline usage. Replications across different elections and countries, with richer direct measures of exposure and belief change, would further clarify causal pathways.

• Generalizability: Estimates apply to participants willing to deactivate for offered payments during the 2020 election; effects may differ in other populations, time periods, or non-election contexts.
• Time-limited intervention: Only 5 weeks of deactivation were studied; longer-term effects may accumulate differently.
• General equilibrium effects: Individual-level deactivation may not capture system-wide changes if many users or entire platforms were reduced; media coverage and information flows could adjust.
• Measurement: Many key outcomes are self-reported, introducing potential measurement error; however, some outcomes are directly measured (platform data, validated turnout, donations, passive tracking).
• Attrition: Endline survey attrition (9–14%, ~2 pp higher in Deactivation) could bias estimates; multiple tests (balance checks, Oster bounds, Ghanem et al. tests, Lee bounds) suggest limited bias, but residual concerns remain.
• Experimenter demand: Despite design features to reduce demand (both groups deactivated at least briefly), participants knew they were in an experiment, which could influence behavior.
• Compliance: Some participants reactivated accounts; IV estimation accounts for imperfect compliance, but residual noncompliance may attenuate or complicate effects.